This paper deals with the approximation of probability density functions with dependent feature trees. The optimal dependent feature trees are proposed to be constructed using criteria of mutual information and distance measures. Expressions are derived for the criteria when the distributions of the features are Gaussian. Expressions are developed for the covariances between the features connected by a path in a dependent feature tree. The case when the nodes in a dependent feature tree represent a set of features is also considered. Furthermore, experimental results from the classification of remotely sensed multispectral scanner imagery data are presented.